# Demographics and Financial Crises: Age Structure as an Early Warning Signal

## 1. Introduction

Financial crises impose enormous economic and social costs, yet their prediction remains elusive. Standard early warning models rely on macroeconomic imbalances — current account deficits, credit booms, fiscal deterioration — but largely ignore the slow-moving demographic forces that shape savings, investment, and external balance patterns. This paper tests whether demographic structure, as captured by the three-factor decomposition of Higgins (1998) and Higgins and Williamson (1997), predicts financial crisis probability, current account reversals, and crisis severity.

Our approach connects two literatures. First, the demographics-capital flows nexus established in our companion papers: demographic structure is a robust predictor of current account balances, with young-dependent and old-dependent populations running deficits and working-age-heavy populations running surpluses. Second, the financial crisis prediction literature, which identifies current account deficits, credit booms, and weak reserves as leading indicators of crises (Kaminsky, Lizondo, and Reinhart 1998; Frankel and Saravelos 2012).

The link between demographics and crisis vulnerability operates through several channels. Youth bulges drive rapid credit growth as expanding working-age populations demand housing and consumer credit, potentially generating lending booms that end in banking crises (Doerr et al. 2022). Aging societies face different risks: shrinking labor forces compress bank net interest margins, incentivizing reach-for-yield behavior that increases systemic fragility. Meanwhile, demographic structure shapes the current account, and large deficits — particularly when financed by volatile capital flows — increase vulnerability to sudden stops.

We merge the Laeven and Valencia (2018) systemic crisis database with our 140-country demographic panel spanning 1950–2024 (237 countries total, 174 with sufficient data for regression). Our main findings are:

1. **Crisis prediction**: Demographic factors improve Brier score from 0.0139 to 0.0138 for banking crisis onset and from 0.135 to 0.133 for CA reversals (logit estimation), with AUC increasing from 0.708 to 0.731 for banking crises and from 0.632 to 0.659 for CA reversals. Banking crisis onset occurs in only ~0.8% of country-years, so even substantial relative improvements translate to modest absolute gains — a 49% improvement in logit pseudo-R² (0.044 to 0.066) that amounts to marginal Brier score reduction. Out-of-sample testing on the 2007–2012 crisis wave confirms this asymmetry: the demographic-augmented model fails to achieve meaningful recall for banking crisis onset, but substantially improves CA reversal prediction (recall increases from 0.06 to 0.27, AUC from 0.606 to 0.625). Demographics are a systematic risk factor for current account crises, not a banking crisis crystal ball.
2. **Sudden stops**: Age structure predicts current account reversals, with early-transition (young) countries significantly more prone to sharp CA deteriorations. Youth dependency raises reversal risk (+0.094***) while old-age dependency is strongly protective (−0.593***).
3. **NFA asymmetry**: NFA creditor status provides substantial protection against sudden stops, consistent with the creditor/debtor asymmetry documented in our multilateral analysis.
4. **Severity**: Conditional on crisis occurrence, aging countries experience significantly milder post-crisis output losses (R² = 0.276), consistent with the low-volatility equilibrium documented in our companion Japanification paper.

## 2. Data

### 2.1 Crisis Episodes

We use the Laeven and Valencia (2018) systemic banking crisis database, which identifies 144 systemic banking crisis episodes matched to our panel. We supplement this with currency and sovereign debt crisis episodes from the same source. Crisis onset is defined as the first year of each episode, while the crisis indicator spans the full episode duration.

### 2.2 Sudden Stops and CA Reversals

Following Calvo (1998), we define a current account reversal as a year-on-year decline in CA/GDP of at least 3 percentage points (1,350 events in the panel). This broad definition captures both involuntary sudden stops and voluntary CA adjustments. We therefore employ three tightened definitions as robustness: (i) *persistent reversals* — $\Delta$CA/GDP $\leq$ −3pp where CA/GDP remains below the prior three-year average for at least two subsequent years (544 events); (ii) *flow-collapse reversals* — $\Delta$CA/GDP $\leq$ −3pp accompanied by a decline in gross capital inflows (531 events); and (iii) *strict sudden stops* — CA reversals accompanied by negative GDP growth (276 events). A stricter 5-percentage-point threshold (816 events) serves as an additional robustness check.

### 2.3 Demographic Variables

Demographic structure is captured by three factors ($Z_1$, $Z_2$, $Z_3$) from a principal components decomposition of age-specific dependency ratios, following Higgins (1998). We also examine youth dependency and old-age dependency ratios separately to identify opposing channels.

### 2.4 Early Warning Variables

Standard early warning controls include: lagged CA/GDP, fiscal balance, reserves-to-liabilities ratio, GDP growth, inflation, capital account openness (KAOPEN), and lagged NFA position. These follow Kaminsky, Lizondo, and Reinhart (1998) and Frankel and Saravelos (2012).

## 3. Crisis Prediction: Do Demographics Signal Vulnerability?

### 3.1 Baseline Results

Given that crisis onset is a rare binary outcome (banking crisis onset = 0.8% of observations, any crisis onset = 2.2%), we estimate pooled logit models as our primary specification, reporting marginal effects at means:

$$
\Pr(\text{Crisis Onset}_{it} = 1) = \Lambda(\beta \cdot Z_{it} + \gamma \cdot X_{it})
$$

where $\Lambda(\cdot)$ is the logistic CDF, $Z_{it}$ contains demographic factors, and $X_{it}$ contains early warning controls. We also estimate complementary log-log (cloglog) models, which are better suited to rare events due to their asymmetric link function, and report LPM (panel GLS with country and year fixed effects) for comparison. Table 27 presents the primary logit results; Tables 1–3 retain LPM results for comparability with prior literature.

Demographics improve in-sample discrimination for banking onset but do not deliver useful out-of-sample recall; for CA reversals, improvements are meaningful in and out of sample. For banking crisis onset, adding $Z_1$, $Z_2$, $Z_3$ to the logit EW model increases pseudo-R² from 0.044 to 0.066 and AUC from 0.708 to 0.731. In absolute terms, the Brier score improves from 0.01392 to 0.01384 — a modest gain reflecting the fundamental difficulty of predicting events that occur in fewer than 1% of country-years. For CA reversals (base rate 7.6%), the improvement is more substantial: Brier score falls from 0.135 to 0.133, and AUC increases from 0.632 to 0.659 (Table 27). Sign and significance patterns are consistent across logit, cloglog, and LPM estimators. Banking calibration above 10% predicted probability is not informative due to sparse mass (fewer than 15 observations total in those bins); we interpret banking model calibration primarily via Brier score and log loss rather than bin-level comparisons.

### 3.2 Youth Bulge vs. Aging

Youth dependency and old-age dependency have opposing effects on crisis risk, consistent with different underlying mechanisms. Young populations drive credit demand and current account deficits, while aging societies face bank profitability pressures. We note that youth_dep and old_dep are strongly negatively correlated ($r$ = −0.71, VIF $\approx$ 2.3), so their individual coefficients in joint specifications should be interpreted cautiously; joint F-tests are the appropriate significance measure (see Section 4.3).

### 3.3 Crisis Type

Table 2 separates banking, currency, and sovereign crisis onsets. Demographic predictors may differ across crisis types: youth bulges associate more with banking and currency crises (credit booms and CA deficits), while the demographic-sovereign crisis link operates through fiscal channels.

### 3.4 Capital Account Openness Interaction

A natural hypothesis is that financial openness amplifies demographic crisis vulnerability: countries with open capital accounts face larger capital flow swings driven by demographic forces, potentially increasing crisis risk. We test this in three ways: continuous Z × KAOPEN interactions, KAOPEN tercile splits, and separate youth/aging channels by openness level (Tables 23–26).

The results reject the amplification hypothesis. The Z × KAOPEN interaction terms are uniformly insignificant for both banking crises and CA reversals. The tercile splits reveal a more nuanced pattern. For banking crises, demographics predict onset with roughly similar magnitude across all three terciles, though no single tercile achieves statistical significance — a power issue given the split samples. For CA reversals, the demographic channel is concentrated in the middle tercile (all three Z factors significant, with $Z_2$ and $Z_3$ at p < 0.01), while fully open economies show no significant demographic effect. The closed tercile shows marginal significance ($Z_1$ at p < 0.10).

The youth and aging channels behave differently across openness levels. The aging → banking risk channel (Doerr et al. 2022) operates in *financially closed* economies: old-age dependency is associated with banking crisis onset only in the lowest KAOPEN tercile (0.227**, p < 0.05), not in open economies. This suggests that aging-related bank risk-taking is most dangerous where limited financial integration prevents risk diversification. Conversely, the protective effect of aging on CA reversals is strongest in *open* economies: old_dep = −0.838*** in the high-KAOPEN tercile, consistent with aging surplus countries using open capital accounts to accumulate external assets that buffer against reversals.

Why doesn't openness amplify demographic risk? The composition of the KAOPEN terciles provides a clue (Table 26). Financially open countries tend to be wealthier, older (higher old_dep), and run larger surpluses — they are structurally less crisis-prone regardless of their demographic position. Closed economies are younger, poorer, and more volatile, but their limited capital account integration also constrains the external financing that makes sudden stops possible. The middle tercile — countries actively opening but without the institutional buffers of advanced economies — is where demographic forces are most destabilizing.

A natural concern is that the middle-openness result is compositional — middle-KAOPEN countries are disproportionately middle-income, and income rather than openness per se may drive vulnerability. Table 30 addresses this by computing KAOPEN terciles *within* GDP per capita quartiles. The within-income tests show that the demographic effect is strongest in the *closed* tercile of upper-middle-income countries (Q3): $Z_2$ = 0.320*** and $Z_3$ = −0.013*** (both p < 0.01), while the middle KAOPEN tercile within Q3 is insignificant. The Q2 middle tercile does show strong effects ($Z_1$ = −4.05***, $Z_2$ = 0.586**), suggesting that demographic destabilization at intermediate openness is concentrated in lower-middle-income countries. The continuous Z × KAOPEN interaction remains insignificant on logit (all p > 0.10), and adding GDP per capita as a control does not alter this conclusion. The middle-openness result is partially compositional — it is strongest where demographic and financial transitions coincide — but not reducible to income alone.

## 4. Sudden Stops and CA Reversals

### 4.1 Demographic Prediction of Reversals

Table 4 tests whether demographic structure predicts current account reversals ($\Delta$CA/GDP $\leq$ −3pp). Since demographics shape the current account level, they should also predict abrupt adjustments when imbalances become unsustainable.

Table 28 compares results across the tightened reversal definitions introduced in Section 2.2. The demographic signal is strongest for *strict sudden stops* — reversals accompanied by negative GDP growth — where all three Z factors are significant at p < 0.01 (logit). Persistent reversals and flow-collapse reversals show weaker Z-factor significance, suggesting that the original 3pp definition captures a mix of involuntary and voluntary adjustments. The age decomposition is more robust: old_dep is protective across all definitions (logit marginal effects), ranging from −0.114* (flow-collapse) to −0.799*** (5pp threshold). We retain the original 3pp definition as the primary measure for comparability with the literature, but report tightened definitions as robustness confirming that the protective effect of aging is not driven by voluntary CA adjustments.

### 4.2 NFA Creditor Protection

The Z × NFA interaction tests whether NFA creditor status protects against sudden stops. Creditor countries can draw down external assets during stress, while debtors face rollover risk.

### 4.3 Opposing Channels

Youth dependency and old-age dependency have opposing effects on reversal risk. Young populations generate deficits that are vulnerable to sudden stops; aging populations transition toward surpluses that reduce reversal vulnerability. When youth_dep and old_dep enter jointly, the correlation between them ($r$ = −0.71) produces VIF values of 2.31 and 2.25 respectively, making individual coefficient interpretation unreliable. The joint F-test is the appropriate test: for CA reversals, the joint significance of youth_dep and old_dep is overwhelming (F = 37.3, p < 0.001), confirming that demographics predict reversals even though the individual coefficients may be imprecisely estimated. For banking crisis onset, the joint test is marginally significant (F = 3.2, p = 0.04), consistent with the Z-factor specification being the more appropriate demographic measure for banking crises. The sign flip on youth_dep observed in joint specifications (Table 6) reflects this collinearity, not a reversal of the underlying youth → crisis vulnerability channel documented in the univariate specification.

### 4.4 Demographic Tercile Heterogeneity

We split the sample by demographic transition stage (early, mid, late) to test whether early-transition countries face systematically higher reversal risk.

### 4.5 Post-Reversal Recovery

Conditional on experiencing a reversal, does demographic structure predict recovery speed? Table 7 examines $\Delta$CA/GDP in years $t+1$ and $t+2$ following a reversal episode.

## 5. Crisis Severity and Type

### 5.1 Cross-Tabulations

NFA creditor/debtor status and demographic transition stage create distinct crisis type profiles. Table 8 cross-tabulates NFA status with crisis type, while Table 9 examines demographic stage.

### 5.2 Aging and Bank Risk

Following Doerr et al. (2022), old-age dependency may increase banking crisis risk through the bank profitability channel. As populations age, deposit growth slows, loan demand shifts from growth-oriented to consumption, and banks face margin compression that incentivizes risk-taking.

### 5.3 Post-Crisis Output Loss

Conditional on banking crisis onset, we examine whether demographic structure predicts the severity of output losses. We employ four severity metrics: (i) cumulative output gap over five years (model-dependent, requires trend estimation, N = 27); (ii) peak-to-trough real GDP decline (model-free, N = 136); (iii) cumulative growth shortfall relative to pre-crisis average growth (model-free, N = 125); and (iv) cumulative growth shortfall relative to world average (model-free, N = 144). Table 29 presents results.

The demographic-severity result is robust to model-free metrics. The growth shortfall measure produces $Z_1$ = −302* (p = 0.058) and $Z_3$ = −2.06** (p = 0.048), confirming that demographics predict crisis depth without relying on trend estimation. The vs-world shortfall yields the strongest age result: old_dep = −151.5*** (p < 0.001). Peak-to-trough GDP decline is null for Z factors but directionally consistent. The cumulative output loss on the full crisis subsample (N = 144) yields R² = 0.276 with all three Z factors significant at p < 0.01, while the output gap specification (R² = 0.155 on 27 episodes) shows Z significance at p < 0.10. This finding connects directly to our companion Japanification paper, which documents that demographics predict a low-growth, low-volatility equilibrium rather than discrete crisis events. The milder output loss in aging economies is not because crises are less damaging per unit of shock; it is because the pre-crisis growth trajectory is already lower, compressing the gap between trend and trough. An economy growing at 1% that contracts to −1% loses less cumulative output than one growing at 5% that contracts to −3%, even though the latter's contraction is shallower in absolute terms.

The Japanification paper identifies a 15% OADR threshold above which economies enter a structurally low-growth regime. Countries above this threshold that experience banking crises are already operating in the low-volatility equilibrium — their crises are "shallow recessions from shallow expansions." The savings glut associated with aging provides additional self-insurance: high domestic savings reduce dependence on external financing, limiting the amplification channel through which sudden stops convert financial stress into deep recessions. This is the benign face of demographic stagnation: less growth, but also less fragility.

### 5.4 Income Group Heterogeneity

The demographic-crisis nexus may differ between high- and low-openness countries. Countries with open capital accounts face larger external adjustment risks but may also have better institutional capacity for crisis management.

## 6. Out-of-Sample Validation

### 6.1 Predicting the 2007–2012 Crisis Wave

We split the sample at 2006 and train the early warning model on pre-crisis data, then evaluate predictions on the 2007–2012 crisis window. Table 15 reports out-of-sample performance for both the standard early warning (EW) model and the EW + demographics (EW + Z) specification. The classification threshold is set to the training-set event base rate (i.e., observations with predicted probability exceeding the in-sample crisis frequency are classified as predicted events). This base-rate threshold is appropriate for surveillance applications where matching the unconditional event frequency avoids the false sense of precision from an arbitrary 0.5 cutoff on rare outcomes.

For banking crisis onset, neither the EW-only nor the EW+Z model achieves meaningful recall — consistent with the well-known difficulty of predicting rare events out of sample. However, for CA reversals, the addition of demographic factors substantially improves prediction: the EW+Z model achieves AUC of 0.625 compared to 0.606 for EW-only, with recall increasing from 0.06 to 0.27. Demographics identify structurally vulnerable countries even when trained on pre-crisis data.

### 6.2 Forward Vulnerability Rankings

Using the full-sample model and 2024 demographic data, we rank countries by predicted crisis vulnerability (Table 16). The ranking combines the systematic demographic component ($X\beta$) with estimated country fixed effects. Countries with prior crisis histories (Ukraine, Argentina, Russia) rank highest, but the demographic component identifies additional vulnerability in countries undergoing rapid demographic transition (Mongolia, Kazakhstan, Turkey).

### 6.3 Demographic Trajectories

Table 17 examines which countries are aging fastest and which retain elevated youth dependency. Small states aside, Puerto Rico, Albania, and Monaco lead in aging speed, while countries like Cape Verde, Bhutan, and Laos are experiencing the fastest youth dependency declines — transitioning from high-reversal-risk to lower-risk demographic profiles.

## 7. Robustness

### 7.1 LPM as Robustness

Our primary specification uses logit (Section 3.1), which is appropriate for rare binary outcomes. Table 18 confirms that LPM (panel GLS with country and year fixed effects) produces qualitatively identical results: $Z_1$ is negative and significant for banking crisis prediction in LPM (−0.159**) and logit (MFX = −0.134**). We additionally estimate complementary log-log (cloglog) models, which have an asymmetric link function better suited to rare events where $\Pr(y=1)$ is small. Cloglog results are virtually identical to logit (Table 27), confirming that the choice between these nonlinear estimators does not drive conclusions. The LPM has the advantage of incorporating country and year fixed effects directly; we report it alongside logit for transparency.

### 7.2 Five-Year Lagged Demographics

Table 19 replaces contemporaneous $Z_t$ with five-year lagged values $Z_{t-5}$ to address potential simultaneity. Since demographic structure is predetermined (births occur decades before working-age participation), this is a conservative test. Results are virtually unchanged: $Z_1$ lagged five years has coefficient −0.169** for banking crisis onset, compared to −0.159** for contemporaneous $Z_1$ (verified in Table 31). This stability confirms that demographics are genuinely predictive rather than reflecting crisis-induced compositional shifts.

### 7.3 OECD vs. Non-OECD

Table 20 splits the sample by OECD membership. For banking crisis prediction, demographic factors are significant in the non-OECD subsample ($Z_1$ = −0.140**, p < 0.05) but not in the OECD subsample ($Z_1$ = −0.422, p > 0.10). The OECD null is not purely a power issue from 37 countries. Three structural features of advanced economies suppress the demographic crisis signal.

First, **institutional backstops**. OECD banking systems operate within deposit insurance frameworks, lender-of-last-resort facilities, and — since 2008 — explicit "too big to fail" guarantees. These backstops transform what would otherwise be a banking crisis into a fiscal or monetary policy event. The demographic pressures that generate banking fragility (margin compression, reach-for-yield) still operate in OECD economies, but they manifest as slow-burn profitability erosion rather than acute crisis onset. Japan's banking sector has experienced precisely this: decades of demographic drag on profitability without triggering a crisis episode that would appear in the Laeven-Valencia database after 1997.

Second, **monetary policy intervention**. Central banks in aging OECD economies have systematically intervened to prevent banking stress from becoming systemic — quantitative easing, yield curve control, negative interest rate policy. These interventions effectively transfer demographic banking risk from the financial system to the central bank's balance sheet, removing the crisis signal from our dependent variable.

Third, **the composition of demographic risk differs — and the direction of the aging effect reverses between OECD and non-OECD**. In non-OECD countries, the demographic-crisis channel operates through *both* youth-driven credit booms *and* aging-driven bank fragility. The aging-banking channel shows this starkly: old-age dependency is *positively* associated with banking crisis risk in non-OECD countries (old_dep = 0.098**, p < 0.05; youth_dep = 0.023**, p < 0.05), meaning that aging *increases* banking crisis risk in developing economies. In OECD countries, old-age dependency is negative but insignificant (−0.080, p > 0.10). This reversal is a key finding of the expanded 140-country panel: aging is not universally protective against banking crises.

However, further decomposition reveals that the non-OECD sign flip is **not an aging phenomenon but a low-income phenomenon**. Splitting the non-OECD subsample by income terciles concentrates the entire effect in the poorest countries:

- **Low income** (GDP/pc < $4,840): old_dep = +0.560** (p = 0.018) — the *entire* positive effect resides here
- **Middle income** ($4,840–$13,678): old_dep = +0.090 (NS, p = 0.152)
- **High income** (> $13,678): old_dep = −0.136 (NS, p = 0.245)

The interaction old_dep × log_gdp_pc = −0.138*** (p = 0.005) confirms this monotonic attenuation: at the lowest GDP per capita levels, the implied old_dep coefficient is approximately +1.50, diminishing toward zero at higher income. Crucially, GDP per capita does *not* absorb old_dep directly — old_dep survives at p = 0.021 with log_gdp_pc as a control — so this is not income masquerading as aging. It is the *interaction* of even modest demographic shifts with weak institutional capacity that generates banking fragility.

Two institutional moderators clarify the mechanism. Foreign exchange reserves buffer the effect: old_dep × reserves = −0.003** (p = 0.036), consistent with reserves providing self-insurance against deposit outflows in systems lacking formal deposit protection. Gross liabilities and trade openness interactions are insignificant, indicating that external balance sheet exposure is not the operative channel. Pension spending data are available for only four non-OECD countries, precluding a direct test of the pension-system mechanism, but the income gradient strongly suggests that the absence of formal old-age support systems — not pension costs per se — is the underlying driver.

Furthermore, the Z polynomial specification is completely null in non-OECD countries (all p > 0.13), while it is strongly significant in OECD countries (all p < 0.01). The non-OECD banking crisis signal comes entirely from the old_dep *level* variable, not from the polynomial shape that captures the lifecycle savings profile. This is consistent with a threshold mechanism: in very poor countries, even modest increases in old-age dependency (from approximately 3% to 6%) strain rudimentary banking systems that lack deposit insurance, pension coverage, and adequate reserve buffers. The relevant demographic range is fundamentally different from OECD aging (old_dep = 0.20–0.35).

The proper characterization is therefore **demographic stress in low-income banking systems**, not "aging causes banking crises in developing countries." In the poorest economies, small increases in the elderly share create deposit withdrawals and loan-quality deterioration in banking systems with no institutional shock absorbers. As countries cross into middle-income status and develop deposit insurance, pension systems, and deeper financial markets, this vulnerability disappears. The Doerr et al. (2022) reach-for-yield mechanism — which requires sophisticated banks responding to margin compression with deliberate risk-taking — is an OECD phenomenon operating through a completely different channel at a completely different demographic scale.

## 8. Discussion and Conclusion

Demographics represent a slow-moving but powerful predictor of financial crisis vulnerability. Our results suggest that standard early warning models omit a systematic risk factor: the age structure of the population shapes savings-investment balances, credit demand, and current account dynamics in ways that affect crisis probability.

Four findings stand out.

**First, demographics improve in-sample discrimination for banking onset but do not deliver useful out-of-sample recall.** For banking crisis onset (base rate 0.8%), the logit pseudo-R² increases by 49% (0.044 to 0.066) and AUC from 0.708 to 0.731, though the absolute Brier score improvement is marginal (0.0139 to 0.0138) and out-of-sample recall is zero — reflecting the fundamental limit of rare-event prediction. For CA reversals (base rate 7.6%), the improvement is meaningful across all metrics: Brier score falls from 0.135 to 0.133, AUC increases from 0.632 to 0.659, and out-of-sample recall jumps from 0.06 to 0.27. Both logit and cloglog estimators confirm the result. Demographics are a systematic surveillance indicator for external adjustment vulnerability, not a banking crisis crystal ball.

**Second, youth dependency and old-age dependency have sharply opposing effects** on reversal risk: young populations face structurally higher vulnerability to sudden stops (+0.094***), while aging populations are increasingly protected (−0.593***). For banking crises, the aging effect is not universally protective: old-age dependency *increases* banking crisis risk in non-OECD economies (0.098**, p < 0.05), while remaining insignificant in OECD countries. However, this sign flip is a **low-income phenomenon**, not an aging phenomenon: income tercile decomposition concentrates the entire positive effect in the poorest non-OECD countries (old_dep = +0.560**, p = 0.018 for GDP/pc < $4,840; insignificant at higher income levels), with a significant income interaction (old_dep × log_gdp_pc = −0.138***, p = 0.005). Even modest demographic shifts strain rudimentary banking systems that lack deposit insurance and formal pension coverage — a fundamentally different mechanism from OECD aging at three to five times the old-age dependency ratio.

**Third, conditional on crisis occurrence, demographic structure predicts crisis severity** across multiple metrics. This result holds on model-free measures — cumulative growth shortfall vs pre-crisis average ($Z_1$ = −302*, $Z_3$ = −2.06**) and growth shortfall vs world average (old_dep = −151.5***) — and is not an artifact of output-gap trend estimation. The full crisis subsample (N = 144) yields R² = 0.276. Aging countries experience significantly milder post-crisis contractions.

**Fourth, middle openness identifies a suggestive vulnerability window.** The KAOPEN tercile analysis (Table 24) reveals that demographic destabilization is concentrated in countries actively integrating into global capital markets but lacking the institutional buffers of advanced economies. For CA reversals, all three Z factors are significant at p < 0.01 in the middle tercile, while neither closed nor fully open economies show a significant demographic effect. Within-income-group tests (Table 30) confirm this is not purely compositional: demographic effects on CA reversals remain significant within narrow income bands, particularly in the lower-middle and upper-middle income quartiles where demographic and financial transitions coincide. This identifies a suggestive vulnerability window. Countries in the middle tercile are precisely those undergoing simultaneous demographic and financial transitions: their working-age populations are expanding (generating credit demand and CA deficits), their capital accounts are opening (enabling external financing of those deficits), but their institutional frameworks — prudential regulation, reserve management, exchange rate flexibility — have not yet matured to buffer the resulting volatility. Mongolia, Kazakhstan, Turkey, and several Southeast Asian economies currently occupy this intersection. For IMF surveillance, this vulnerability window implies that the standard advice to liberalize capital accounts should be conditioned on demographic stage: countries with rapidly expanding working-age populations face elevated reversal risk from premature opening, while aging surplus economies can open safely because their demographic profile generates surpluses rather than deficits.

### Policy Implications

**Crisis prevention.** Crisis prevention frameworks should incorporate demographic projections alongside traditional macroeconomic indicators. Countries entering the rapid credit-growth phase of demographic transition — with expanding working-age populations and declining youth dependency — face elevated banking crisis risk that is predictable decades in advance. The middle-openness vulnerability window provides a suggestive screening criterion: countries in the middle KAOPEN tercile with below-median OADR should receive enhanced surveillance for CA reversal risk.

**NFA accumulation as dual-purpose insurance.** The NFA creditor/debtor asymmetry implies that building external buffers during the demographic surplus phase serves a dual purpose. Our multilateral paper documents that NFA creditor status produces a CA coefficient of +0.73 (p = 0.0002) while debtor status yields −0.14 (p = 0.36) — creditors actively manage their external positions while debtors are passive. In the crisis context, this asymmetry translates directly into crisis resilience: creditor countries can draw down external assets during stress rather than facing rollover risk on external liabilities. The policy prescription is therefore not merely to "save during the surplus phase" but to accumulate *external* assets specifically — sovereign wealth funds, reserve accumulation, pension fund foreign allocation — that provide both income (the current account income balance channel) and sudden-stop insurance simultaneously. Countries that run surpluses but hold them as domestic assets (bank deposits, government bonds) miss the crisis insurance benefit.

**Forward vulnerability surveillance.** Table 16 provides a demographic vulnerability heat map using 2024 data. While countries with prior crisis histories (Ukraine, Argentina, Russia) rank highest due to estimated country fixed effects, the demographic component identifies structurally vulnerable countries that standard macro indicators may miss. Mongolia ranks high because its rapid demographic transition — one of the fastest-declining youth dependency ratios in Asia — is generating a credit boom as its expanding working-age population demands housing and consumer finance, while its capital account sits in the volatile middle tercile and its commodity-dependent economy amplifies external shocks. Turkey's vulnerability has a different demographic signature: its working-age bulge coincides with an already-open capital account and chronic CA deficits, placing it in the "high demographic pressure, high financial exposure" quadrant. Both countries illustrate the model's value: their macro profiles differ substantially (commodity exporter vs. manufacturing/services, small vs. large economy), but their demographic positions generate similar structural vulnerability to CA reversals.

These findings complement our companion papers on demographics and capital flows. The multilateral analysis establishes that demographics drive current account patterns; the bilateral gravity model traces the geographic direction of these flows; the causal identification paper confirms the structural nature of the demographic-CA relationship. This paper closes the loop by showing that demographic imbalances not only predict capital flows but also predict when those flows become destabilizing.

## References

Calvo, G.A. (1998). Capital flows and capital-market crises: the simple economics of sudden stops. *Journal of Applied Economics*, 1(1), 35–54.

Doerr, S., Kabas, G., and Ongena, S. (2022). Population aging and bank risk-taking. *Journal of Financial Economics*, 146(2), 353–378.

Frankel, J. and Saravelos, G. (2012). Can leading indicators assess country vulnerability? Evidence from the 2008–09 global financial crisis. *Journal of International Economics*, 87(2), 216–231.

Higgins, M. (1998). Demography, national savings, and international capital flows. *International Economic Review*, 39(2), 343–369.

Higgins, M. and Williamson, J.G. (1997). Age structure dynamics in Asia and dependence on foreign capital. *Population and Development Review*, 23(2), 261–293.

Kaminsky, G.L., Lizondo, S., and Reinhart, C.M. (1998). Leading indicators of currency crises. *IMF Staff Papers*, 45(1), 1–48.

Laeven, L. and Valencia, F. (2018). Systemic banking crises revisited. *IMF Working Paper* WP/18/206.
